Skip to content

Latest commit

 

History

History
241 lines (168 loc) · 6.6 KB

README.md

File metadata and controls

241 lines (168 loc) · 6.6 KB

Concurrency and Parallelism (Node.js, Worker Threads, Shared Memory)

This module is about correctness under concurrency pressure, not toy async syntax.

Scope:

  • What JavaScript's "single-threaded" claim means and where it misleads
  • Worker Threads and message passing
  • SharedArrayBuffer + Atomics for shared-memory coordination
  • Deterministic race patterns
  • Thread-safe abstraction design

1) Single-threaded model myths

JS is single-threaded per execution context, not per process

A single JS event loop (one thread) executes one callback at a time in that context. A Node.js process can still run multiple threads:

  • libuv thread pool (for some I/O and native tasks)
  • worker threads (node:worker_threads)
  • OS/network work outside JS thread

Concurrency vs parallelism

  • Concurrency: multiple tasks in progress, interleaved.
  • Parallelism: tasks executing simultaneously on different cores/threads.

Event loop gives concurrency but usually not CPU parallelism for JS code on one thread. Worker threads provide actual JS parallelism.

Where races exist in JavaScript

Races can happen in two major forms:

  1. Logical races on one thread (async interleavings, stale reads).
  2. Data races with shared memory (SharedArrayBuffer) if non-atomic operations are used.

2) Worker Threads (Node)

Message passing model

Workers do not share normal JS objects by default. They communicate through postMessage and structured clone.

'use strict';
const { Worker } = require('node:worker_threads');

const worker = new Worker(`
  const { parentPort } = require('node:worker_threads');
  parentPort.on('message', (n) => parentPort.postMessage(n * 2));
`, { eval: true });

worker.once('message', (result) => {
  console.log(result); // 42
});
worker.postMessage(21);

Structured cloning and transferables

  • postMessage clones data (structured clone).
  • Transferables move ownership (e.g. ArrayBuffer) instead of cloning.
  • SharedArrayBuffer is shared, not cloned.

Lifecycle and cost

Worker creation is not free:

  • startup latency
  • memory overhead
  • serialization cost for messages

Use workers when:

  • CPU-bound tasks are heavy enough
  • partitioning work is clear

Avoid workers when:

  • tasks are tiny
  • overhead dominates
  • shared-state complexity risks correctness

3) SharedArrayBuffer and Atomics

Shared memory model

SharedArrayBuffer lets multiple threads see the same raw bytes. Use typed arrays for views:

'use strict';
const sab = new SharedArrayBuffer(Int32Array.BYTES_PER_ELEMENT * 2);
const view = new Int32Array(sab);

Why Atomics matter

Non-atomic read/modify/write is race-prone:

// Racy pattern (do not use across threads):
view[0] = view[0] + 1;

Atomic increment:

Atomics.add(view, 0, 1);

Core atomic operations

  • Atomics.load/store
  • Atomics.add/sub/and/or/xor
  • Atomics.compareExchange
  • Atomics.wait / Atomics.notify (Int32Array only)

Memory ordering (practical)

Atomics provide synchronization guarantees so writes become visible in coordinated order. Without Atomics, concurrent visibility/order assumptions are unsafe.

Spinlocks vs wait/notify

Spinlock:

  • busy loop with compareExchange
  • burns CPU while waiting

Wait/notify:

  • block with Atomics.wait
  • wake with Atomics.notify
  • avoids hot spinning under contention

4) Race conditions in JavaScript

Logical race without shared memory (lost update)

'use strict';

let count = 0;

async function inc() {
  const snapshot = count;
  await Promise.resolve(); // interleaving point
  count = snapshot + 1;
}

await Promise.all([inc(), inc()]);
// count can be 1, not 2

Check-then-act bug

if (!cache.has(key)) {
  const value = await load(key);
  cache.set(key, value);
}

Two callers can both miss and both load.

ABA problem (conceptual)

Thread T1 reads A. Thread T2 changes A -> B -> A. T1 sees A again and assumes no change. Value equality alone misses intermediate mutation; version tagging fixes this.

5) Designing thread-safe abstractions

Prefer immutable data + message passing

Most robust default:

  • workers send immutable messages
  • single owner mutates state

Lock-based designs

  • critical sections with mutex/lock
  • clear ownership boundaries
  • ensure unlock in finally-like paths

Lock-free designs (conceptual)

  • CAS loops (compareExchange)
  • version/tag fields to avoid ABA
  • harder to reason about; require strict invariants

Idempotency and commutativity

In distributed/concurrent systems, operations that can be retried or reordered safely reduce race impact.

Backpressure in multi-threaded systems

If producers outrun consumers, buffers grow and memory pressure rises. Bound queues + explicit refusal (push -> false) + retry policy are mandatory.

6) Interview mental model

How to explain JS concurrency clearly:

  1. One event loop is single-threaded.
  2. Process/runtime still uses parallel subsystems.
  3. Worker threads run JS in parallel with explicit communication.
  4. Shared memory requires Atomics for correctness.

What Atomics guarantee:

  • atomicity for specific operations
  • synchronization/visibility guarantees across threads

What Atomics do not guarantee:

  • automatic fairness
  • deadlock freedom
  • correctness of your protocol

Why "JS is single-threaded so no race conditions" is wrong:

  • async interleavings cause logical races on one thread
  • worker + shared memory causes true data races without Atomics

7) Common incorrect assumptions

  • "Await makes operations sequential globally."
  • "Workers can mutate outer-scope variables directly."
  • "SharedArrayBuffer reads/writes are safe without Atomics if values are small."
  • "If code passes tests once, race is solved."
  • "Spinlocks are fine in JS because loops are cheap."

8) Deterministic concurrency testing patterns

For interview-style exercises, avoid timing randomness. Use:

  • explicit barriers (promises, Atomics gates)
  • fixed operation counts
  • deterministic orchestration of interleavings
  • invariant checks (counts, no overlap, no missing items)

Practical constraints & interview caveats

  • Browsers vs Node: SharedArrayBuffer is available in browsers only under cross-origin isolation (COOP/COEP). In Node.js it is available without that requirement.
  • Atomics.wait / Atomics.notify: Atomics.wait only works on Int32Array (and BigInt64Array in newer runtimes) and blocks the calling agent (thread). In Node, it is usable inside Worker Threads.
  • Determinism: Real data races are nondeterministic. Many exercises here use deterministic interleavings to prove correctness of synchronization logic without relying on timing.